Active Fire Detection Using Landsat-8/OLI Data

Remote Sensing of Environment 185 (2016) 210–220

Contents lists available at ScienceDirect

Remote Sensing of Environment

journal homepage: www.elsevier.com/locate/rse

Active ﬁre detection using Landsat-8/OLI data

Wilfrid Schroeder a,⁎, Patricia Oliva a,LouisGiglioa, Brad Quayle b, Eckehard Lorenz c, Fabiano Morelli d a Department of Geographical Sciences, University of Maryland, College Park, MD, USA b USDA Forest Service Remote Sensing Applications Center, Salt Lake City, UT, USA c Institute of Optical Sensor Systems, German Aerospace Center, Berlin, Germany d Brazilian Institute for Space Research, São José dos Campos, Brazil article info abstract

Article history: The gradual increase in Landsat-class data availability creates new opportunities for fire science and management Received 6 April 2015 applications that require higher-fidelity information about biomass burning, improving upon existing coarser Received in revised form 10 August 2015 spatial resolution (≥1 km) satellite active fire data sets. Targeting those enhanced capabilities we describe an Accepted 31 August 2015 active fire detection algorithm for use with Landsat-8 Operational Land Imager (OLI) daytime and nighttime Available online 9 September 2015 data. The approach builds on the fire-sensitive short-wave infrared channel 7 complemented by visible and near-infrared channel 1–6 data (daytime only), while also expanding on the use of multi-temporal analysis to improve pixel classification results. Despite frequent saturation of OLI's fire-affected pixels, which includes radiometric artifacts resulting from folding of digital numbers, our initial assessment based on visual image analysis indicated high algorithm fidelity across a wide range of biomass burning scenarios, gas flares and active volcanoes. Additional field data verification confirmed the sensor's and algorithm's ability to resolve fires of significantly small areas compared to current operational satellite fire products. Commission errors were greatly reduced with the addition of multi-temporal analysis tests applied to co-located pixels, averaging less than 0.2% globally. Because of its overall quality, Landsat-8/OLI active fire data could become part of a network of emerging earth observation systems providing enhanced spatial and temporal coverage of biomass burning at global scales. © 2015 The Authors. Published by Elsevier Inc. This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/).

1. Introduction observed in fire-adapted biomes that benefit from nutrient cycling and seeding (Freeman et al., 2007; Higgins et al., 2007; Oliveras et al., Each year hundreds of thousands of biomass burning related fires 2012). In areas such as the western United States, southern Australia, are detected globally using spaceborne remote sensing data (Dwyer, and the Iberian Peninsula, large wildfires occurring along the wildland- Pinnock, Grégoire, & Pereira, 2000; Giglio, Csiszar, & Justice, 2006; urban interface pose risks to both life and property and are cause of Ichoku, Giglio, Wooster, & Remer, 2008). Satellite-detected thermal- major socioeconomic concern (Mell, Manzello, Maranghides, Butry, & anomalies are predominantly associated with land use practices Rehm, 2010). Management efforts associated with those events typically (maintenance and conversion fires), wildfires ignited by lightning or involve large investments in fuel treatments, community response plan- human causes and other natural occurring phenomena or processes ning, active fire mapping and suppression, and post-fire stabilization and (e.g., volcanic activity, etc.). Industrial heat sources can also be detected restoration (Schoennagel, Nelson, Theobald, Carnwath, & Chapman, (Bowman et al., 2009; Schroeder et al., 2008b). Biomass burning effects 2009). can be observed across local to global scales, impacting soil chemistry, For more than a decade, satellite remote-sensing active fire data surface runoff, land surface heat and energy balances, and air quality, have been extensively used to inform fire management systems among others (Larsen et al., 2009; Liu, Randerson, Lindfors, & Chapin, (Davies, Ilavajhala, Wong, & Justice, 2009). Similarly, numerous air- 2005; Wiedinmyer et al., 2006). Negative ecological effects may result quality and carbon emissions mapping methodologies have benefited from biomass burning in non-fire-adapted vegetation or as a conse- from point source information provided by satellite active fire detection quence of altered fire regimes (e.g., encroachment of invasive species and characterization data sets (Ichoku & Kaufman, 2005; Kaiser et al., and changes to stand structure), whereas positive effects are typically 2012; van der Werf et al., 2010; Vermote et al., 2009). Other related active fire data applications include assessment of fire-affected areas and seeding of burned area algorithms (Giglio, Loboda, Roy, Quayle, & Justice, 2009; Hantson, Padilla, Corti, & Chuvieco, 2013; Loboda, O'Neal ⁎ Corresponding author at: Department of Geographical Sciences, University of Maryland, 2181 LeFrak Hall, College Park, MD 20742, USA. & Csizar 2007; Kasischke, Hewson, Stocks, van der Werf, & Randerson, E-mail address: [email protected] (W. Schroeder). 2003; Oliva & Schroeder, 2015), fire growth and spread rate analyses

http://dx.doi.org/10.1016/j.rse.2015.08.032 0034-4257/© 2015 The Authors. Published by Elsevier Inc. This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/). W. Schroeder et al. / Remote Sensing of Environment 185 (2016) 210–220 211

(Csiszar & Schroeder, 2008; Loboda & Csizar, 2007; Pozo, Olmo, & (WRS-2) of path and row coordinates (Irons et al., 2012). Quantized Alados-Arboledas, 1997), and detection of fossil fuel emissions sources and calibrated scaled digital numbers (DNs) for each OLI band are deliv- from gas flares (Casadio, Arino, & Serpe, 2012; Elvidge et al., 2009; ered as 16-bit unsigned integers. Those are converted to top-of- Elvidge, Zhizhin, Hsu, & Baugh, 2013), among others. atmosphere (TOA) spectral radiance and planetary reflectance values The use of Landsat-class data to detect thermal anomalies has been using the rescaling coefficients found in the metadata (MTL) file successfully demonstrated in previous studies. For example, Francis available with the L1T data. and Rothery (1987) and Oppenheimer (1991) applied near infrared The Landsat-8 active fire detection builds on previous algorithms de- (NIR) and short-wave infrared (SWIR) Landsat-5 Thematic Mapper veloped for ASTER and Landsat-7 ETM+(Giglio et al., 2008; Schroeder data to case study analyses of volcanic activity. Others expanded on et al., 2008a). Both methodologies used a two-channel fixed-threshold thermal anomalies applications to include long-wave infrared data plus contextual approach exploring the differential radiometric re- from Landsat-5 Thematic Mapper (TM) and Landsat-7 Enhanced sponse of the SWIR (channel 8 on ASTER; channel 7 on ETM+) and Thematic Mapper Plus (ETM+) (Flynn, Harris, & Wright, 2001; the NIR (channel 3 N on ASTER; channel 4 on ETM+) data to classify Anejionu, Blackburn, & Whyatt, 2014). A more generic application of fire-affected pixels. Here, we expand on that original methodology NIR and SWIR Landsat-class data to detect actively burning fires was using top-of-atmosphere spectral reflectance data from seven different also demonstrated (Giglio et al., 2008; Schroeder et al., 2008a). However, OLI channels. The list of OLI channels used by the fire algorithm is shown because of limited data availability and access restrictions previous stud- on Table 1, along with their intended application. Complementing the ies using those active fire data were mainly focused on regional validation input OLI data described on Table 1, pixel-based cloud coverage infor- analyses of active fire products derived from the 1-km Moderate Resolu- mation is derived from the auxiliary quality band (cloud confidence tion Imaging Spectroradiometer (MODIS) and the 4-km Geostationary bits 14–15) for use in the multi-temporal analysis discussed in Section 4. Operational Environmental Satellite (GOES) imager (Csiszar, Morisette, As with ASTER and ETM+ data, active-fire-affected pixels may often & Giglio, 2006; Morisette, Giglio, Csiszar, & Justice, 2005; Schroeder exceed the maximum resolvable radiance on OLI's fire-sensitive SWIR et al., 2008a). channel 7, causing frequent pixel saturation. Spurious DN values and ar- While the spatial and temporal coverage provided by individual tificial radiances may be output by the analog-to-digital converter when Landsat-class sensors remained relatively unchanged over the years, the input signal exceeds the nominal detector range. Under more ex- the adoption of free data policies and the launch of new instruments treme conditions when large and/or very high temperature fires occupy by international agencies helped gradually increase the availability of the pixel footprint the increased radiance input to the detector can also Landsat-class data creating renewed opportunities for biomass burning lead to folding of the pixel's DN (also known as oversaturation (Morfitt applications. The current United States Geological Survey (USGS) et al., 2015)). This less frequent anomaly is typically characterized by Landsat, China-Brazil Earth Resources Satellite (CBERS), Indian pixels of abnormally low OLI channel 7 radiance values located near Resourcesat, and European Space Agency Sentinel-2 programs are the core of actively burning fire perimeters, an artifact resulting from good examples of the growing number of Landsat-class assets serving the DN folding, accompanied by elevated radiance on the shorter wave- the broader user community. Compared to traditional ≈1-km spatial length channels 6 and 5. Fig. 1 shows a manifestation of the DN folding resolution satellite fire products, Landsat-class data offer significantly on channel 7 coinciding with the large and energetic Rim fire complex improved mapping capability generating detailed fire line information. in California/U.S., imaged on 31 August 2013. A small subset of the fire Those assets have the potential to transform the way satellite data are perimeter on channel 6 reveals a 15 pixel wide area of elevated radi- used in support of fire management, adding to current airborne tactical ances near the core, whereas the co-located channel 7 data show a fire mapping resources, for example, and providing consistent and reli- ring of elevated radiances around an anomalous low-radiance core able fire information to decision support systems operating at similar indicative of DN folding. The graphs on Fig. 1 show the channel 6 and spatial scales. A direct application of such concept was demonstrated 7 radiance profiles drawn across the active fire, and the histogram for by Coen and Schroeder (2013, 2015), who successfully used spatially re- channel 7 radiances coincident with a larger sample consisting of over fined satellite fire data to initialize and later verify coupled weather-fire 4500 fire-affected pixels on the scene. The channel 6 profile shows model simulations of large long-duration wildfires in the western high radiance values near the center of the fire, including pixels at or United States. above the nominal saturation of 71.3 W/(m2 sr μm). The same core In this study, we present an active fire detection algorithm for use area shows near-zero and artificially low channel 7 radiance values with the Landsat-8 day and nighttime data. The approach builds on pre- surrounded by other pixels at or above the nominal saturation of vious algorithms applied to the Advanced Spaceborne Thermal Emis- 24.3 W/(m2 sr μm), therefore describing a typical DN folding scenario. sion and Reflection Radiometer (ASTER) and Landsat-7 ETM+ NIR We note that the maximum observable radiance on channels 6 and 7 and SWIR data while further expanding the use of multi-temporal anal- are 96 and 29 W/(m2sr μm), respectively (Morfitt et al., 2015). ysis to improve the classification of individual pixels. The histogram in Fig. 1 shows the highest number of pixels associated with the Rim fire with radiance values around the nominal saturation 2. Input Data on channel 7. In addition, a non-negligible number of pixels with radiance values above the nominal saturation are found as a result of analog Landsat-8 was designed by the National Aeronautics and Space high saturation. A few pixels corresponding to DN folding are found Administration (NASA) and launched on 11 February 2013 carrying the Operational Land Imager (OLI) and the Thermal Infrared Sensor Table 1 (TIRS), and subsequently transferred to USGS for routine operations List of 30 m resolution Landsat-8/OLI channels used in the active fire detection algorithm, (Roy et al., 2014). Placed on a sun-synchronous orbit at 705 km altitude and their primary application. and 10:00 a.m. equatorial crossing time for the descending node, OLI channel Wavelength Application Landsat-8 has a 16-day repeat cycle although data acquisition strategy (μm) may vary based on geographic coverage (including seasonal sampling – fi and cloud coverage) and overall science mission objectives (Irons, 1 0.43 0.45 Active re detection & water mask 2 0.45–0.51 Water mask Dwyer, & Barsi, 2012; Roy et al., 2014). In this study, we used standard 3 0.53–0.59 Water mask terrain corrected (Level 1 T) data from OLI, which is a nine spectral 4 0.64–0.67 Water mask channel push-broom sensor with a spatial resolution of 30 m (15 m 5 0.85–0.88 Active fire detection & water mask for the panchromatic channel 8), and an individual scene size of 6 1.57–1.65 Active fire detection & water mask 7 2.11–2.29 Active fire detection, water mask & temporal analysis 185 km × 180 km matching the second World-wide Reference System 212 W. Schroeder et al. / Remote Sensing of Environment 185 (2016) 210–220

Fig. 1. Landsat-8/OLI channel 6 (top-left) and 7 (top-center) image subsets acquired on 31 August 2013 over the Rim fire in California/U.S (37.760°N 119. 965°W; path/row 43/34). The low radiance fire core on channel 7 is indicative of DN folding. The radiance profiles (bottom-left, bottom-right) correspond to the horizontal line across each subset. The spurious drop in channel 7 radiance is the result of DN folding; the same pixels show saturated radiances on channel 6. The histogram describing channel 7 radiances for all 4500 fire-affected pixels (marked in red) in the top-right image subset is shown on the bottom-right panel. The peak in the radiance distribution coincides with the nominal saturation radiance (24.3 W/(m2srμm)). Pixels exceeding the nominal saturation are representative of analog high saturation. A small number of pixels describing DN folding can be found near the low end of the radiance range. The small yellow box in the top-right panel indicates the area subject to DN folding.

near the low end of the radiance range on channel 7. While this partic- fires in the 2.2 μm spectral window. During the daytime orbits the emis- ular case describes characteristics associated with significant fire inten- sive fire component is mixed with the background, which is dominated sity for a major wildfire event, folding of OLI channel 7 DN can also be by the reflected solar component. In order to separate those, we use the observed over some large gas flares. NIR channel 5 data that are mostly unresponsive to fire-affected pixels, though highly correlated to the SWIR channel data over fire-free sur- 3. Active fire algorithm faces (Giglio et al., 2008). During night orbits the reflected solar component is absent from the scene, making the SWIR band particularly The Landsat-8 active fire detection algorithm is divided into day and responsive to the emitted radiance from active fires in an otherwise nighttime modules. Both detection modules are driven by the fire- dull background. In both day and nighttime data, the radiometric signa- sensitive SWIR channel 7 data, exploiting the emissive component of ture of active fires produces a SWIR radiance or reflectance anomaly

Fig. 2. Geographic distribution of primary Landsat-8/OLI Level 1 T scenes used for training and validation of the active fire detection algorithm. W. Schroeder et al. / Remote Sensing of Environment 185 (2016) 210–220 213 when compared to the background, thereby mimicking the concept of order to map oceans and inland water bodies according to: thermal anomaly detection using mid-to-thermal infrared channels. Fig. 2 shows the distribution of Landsat-8 scenes used to train and fi ρ Nρ ρ Nρ ρ Nρ ρ −ρ b : later validate the re algorithm output. Scene selection allowed sam- fg4 5 AND 5 6 AND 6 7 AND 1 7 0 2 ð7Þ pling of a wide range of fire and observation conditions. Co-located AND scenes indicate areas where multi-temporal data acquisitions were ρ Nρ obtained, with training and validation data analyzed separately. The fðÞ3 2 ð8Þ training data were used to calibrate the algorithm and were based primarily on histogram analysis of single channel and dual-channel OR (e.g., band ratios and differencing) data supported by detailed super- ρ Nρ ρ Nρ ρ Nρ : vised pixel classification information. The latter was performed by ðÞg1 2 AND 2 3 AND 3 4 ð9Þ expert image analysts and corroborated by high spatial resolution imagery (e.g., aerial photography and commercial satellite imagery Tests (7 & 8) capture shallow and/or sediment-rich waters available in Google Earth) and other remotely sensed fire products. (e.g., Amazon river mouth) whereas tests (7 & 9) are useful for mapping These supporting data were used to validate fire activity identified by deep and/or dark water bodies. This water-masking scheme resulted in the detection algorithm. The algorithm thresholds were defined for a more liberal classification compared to the available water confidence TOA reflectance data that have not been corrected for solar elevation. bits in the L1T quality band, allowing more water pixels to be flagged al- beit with some errors of commission. For example, tests (7 & 9) may 3.1. Daytime detection confuse water bodies and cloud shadows and thereby affect the valid background statistics, although such classification errors had no notice- The daytime algorithm uses input data from all seven OLI channels able effect on the fire algorithm performance. listed in Table 1.Thefirst test in the daytime module is designed to identify potentially unambiguous active fire pixels. It builds on the ETM+ 3.2. Nighttime detection active fire algorithm (Schroeder et al., 2008a) while accommodating small differences in OLI spectral channels, and is based on the following The absence of the reflected solar component from the nighttime condition: scenes makes the classification of fire-affected pixels significantly less challenging. Here we use a single fixed-threshold test based on the N : ρ −ρ N : ρ N : fgðR75 2 5AND 7 5 0 3AND 7 0 5 1Þ SWIR channel 7 radiance data and similar to Giglio et al. (2008): ÀÁ 2 L7 N1W= m sr μm ð10Þ where ρi is the reflectance on channel i,andRij is the ratio between channel i and j reflectances (i.e.,ρi/ρj). As described in Section 2, highly energetic and extensive fires can Where L7 is the channel 7 radiance. We did not encounter evidence lead to folding of the DN on channel 7 thereby characterizing another of DN folding in the nighttime scenes analyzed. We recognize, however, condition of potentially unambiguous active fire pixel. Those unique oc- that our assessment was limited by the reduced availability of nighttime currences are flagged using: scene acquisitions. Additional tests using channel 6 radiance data may be required to properly handle DN folding in those data. ρ N : ρ b : ρ N : ρ b : : fg6 0 8AND 1 0 2ANDðÞ5 0 4OR 7 0 1 ð2Þ 4. Multi-temporal analysis Complementing the identification of unambiguous fire pixels, the thresholds in test (1) are relaxed and other candidate fire pixels are se- The Level 1T OLI data have a geolocation error requirement of 12 m lected for further analyses based on the following criteria: (Irons et al., 2012), greatly improving the co-location of surface features observed on images acquired over multiple dates. Building on the en- N : ρ −ρ N : : fi fgR75 1 8AND 7 5 0 17 ð3Þ hanced geolocation data, we designed a simple procedure to help re ne the fire algorithm output using co-located data derived from previous All pixels satisfying test (3) must then meet the following set of fixed images acquired no more than 176 days apart. This temporal constrain threshold and contextual tests in order to be classified as potential fire- was applied in order to minimize the effects of seasonal variations affected pixels: that could potentially alter the land surface conditions, while balancing cloud-free data availability. ÂÃ The approach is based on direct verification of previously processed R75 NR75 ̅þ max 3σ R ; 0:8 ð4Þ 75 active fire data for spatially-coincident and temporally-persistent fire pixels. This simple test is based on the assumption that typical biomass AND burning fuels within a 30 m ground footprint are completely consumed ÂÃ within minutes, hours, or a few days of continuous flaming and/or smol- ρ Nρ ̅þ max 3σ ρ ; 0:08 ð5Þ 7 7 7 dering combustion. Therefore pixels classified as potential fire-affected areas by the detection algorithm that overlap with one or more previ- AND ously detected fire pixels are assigned a persistent source class indicative of temporally persisting detection. Such classification was found useful N : R76 1 6 ð6Þ to flag gas flares and other stationary urban heat sources (e.g., steel mills). Consequently, it can be used to differentiate between open ̅ σ ρ ̅ σρ fi where Rij and Rij ( 7 and 7) are the mean and standard deviation vegetation res and other types of combustion or heat sources. Addi- calculated for the band ratio (channel 7 reflectance) using valid back- tionally, this test may flag potential daytime commission errors (false ground pixels from a 61 × 61 window centered on the candidate alarms) associated with radiometrically bright urban features (e.g., hot pixel. Valid pixels are defined as those showing channel 7 reflectance and reflective factory rooftops) and other unique structures such as greater than zero, excluding water and unambiguous fire pixels. solar farms/photovoltaic stations, etc. We note that charcoal pits and Water pixels are classified based on spectral profiling using reflectance repeated piling of biomass for burning in designated areas may also be data from all seven input channels. Two distinct tests are applied in flagged as persistent sources as a result of this test. 214 W. Schroeder et al. / Remote Sensing of Environment 185 (2016) 210–220

Pixels output by the daytime fire algorithm without a co-located assess individual algorithms, and to compare between different fire detection in the previous 176 days are further inspected to check for products. the occurrence of high channel 7 reflectances persisting over time, an Here we calculated the fire algorithm's day and nighttime detection indication of potential false alarm. We used channel 7 for this test due curves separately. The daytime fire detection curve calculation built on to its strengths in separating vegetated and non-vegetated (e.g., bare the approach of Schroeder, Oliva, Giglio, and Csiszar (2014),who soil) surfaces, and reduced contamination by smoke (Asner & Lobell, assessed the performance of the Suomi National Polar-orbiting Partner- 2000; Kaufman et al., 1997). The test assigns a bright surface class to ship Visible Infrared Imaging Radiometer Suite (S-NPP/VIIRS) 375-m ac- ρ N : fl fi those pixels with 7;t 0 2, which describes the mean channel 7 re ec- tive re detection algorithm using actual data. We selected 12 different tance calculated using t co-located cloud-free pixels in the previous OLI images acquired globally, each representing a distinct geographic 176 days. This additional test is designed to flag pixels in areas dominat- area at times coinciding with their respective active fire seasons. Fire ac- ed by urban environment or exposed soils that, under specific illumina- tivity varied among individual scenes, ranging from as few as 79 pixels tion conditions, can mimic the biomass-burning signature causing a (Northern Territory/Australia) to a maximum of 6561 pixels (Central false alarm. Those targets may also be found in association with pixels Siberian Plateau/Russia). For each scene, a total of 25 pixels were select- adjacent to persistent urban detections, where the increased distance ed across the entire domain. Pixel selection included areas adjacent to to the heat/high reflectivity source results in less frequent manifestation active fires (simulating an extended fire line), clouds and water bodies, of reflectance anomalies. subject to smoke plumes, and across a wide range of vegetated and non- The application of the multi-temporal analysis component is dem- vegetated areas (randomly sampled). Fires were then simulated on onstrated in Fig. 3, which shows an OLI image subset over eastern those individual pixels by incorporating fire-emitted radiance to OLI's China (path/row 119/41). Potential false alarms associated with radio- primary channels 5, 6, and 7. Simulated pixel radiances were converted metrically bright urban structures can be seen on the fire algorithm back to scaled DN, and then to reflectance using the corresponding MTL output. Detailed visual inspection of those locations revealed factories file parameters. Atmospheric attenuation effects were estimated using a and other facilities consisting of large reflective rooftops and heat ex- radiative transfer code (MODTRAN®) running on atmospheric profiles hausts. Application of the complementary multi-temporal analysis derived from the National Oceanic and Atmospheric Administration tests to those pixels resulted in successful re-classification of all pixel (NOAA) Global Forecast System (GFS) model analysis data for each locations, setting them apart from actual biomass burning activity in scene selected. Fire effective areas were varied between 1 and 150 m2 neighboring rural areas. (1 m2 intervals), and mean temperatures were varied between 400 and 1200 K (10 K intervals). Simulated single-component fire radiances were calculated with consideration to the fractional area coverage and 5. Theoretical fire detection envelope the spectral response function of each channel, and then added to the selected pixel radiances. In order to ensure optimum representation of Satellite fire data users are fundamentally interested in the resolving observable conditions, all background pixels were preserved resulting power of new algorithms, in other words, what minimum fire size can in realistic fire-affected synthetic images. These synthetic data were be typically detected? Effective quantification of satellite fire detection processed using the daytime fire algorithm without any modification. performance using ground truth data can be challenging due to lack of Fig. 4 shows the 50% probability of fire detection curve derived from adequate reference data and, most importantly, because of the high all 3.6 million data points resulting from the simulation scenarios costs involved. Alternatively, theoretical calculation of detection curves above. The curve shows a noticeable increase in fire detection per- as a function of fire size and temperature provides useful information to formance compared to other coarser spatial resolution fire data

Fig. 3. Subset of 07 August 2014 Landsat-8/OLI image of eastern China (path/row 119/41). Green shades indicate vegetation, gray and magenta shades indicate urban areas. Left panel shows pixels output by the fire algorithm; right panel shows the corresponding multi-temporal analysis result with all pixels re-classified as urban-related heat/highly reflective sources. W. Schroeder et al. / Remote Sensing of Environment 185 (2016) 210–220 215

VIIRS active fire data (Csiszar et al., 2014; Giglio, Descloitres, Justice, & Kaufman, 2003; Schroeder et al., 2014). Those satellite images were acquired sequentially, averaging 1-h separation between consecutive over- passes. Overall, the different products showed good correspondence both in terms of fire front location and extent. Notably though, Landsat-8 fire pixels provided a significantly more detailed mapping of the flaming front describing areas of continuous activity on both east and west flanks of the fire, plus isolated heat islands inside the perimeter. The formation of an optically thick pyrocumulus cloud at the northern section (leading edge) of the fire perimeter had variable effects on the fire products as it evolved throughout the day. Other areas dominated by relatively dry smoke showed no noticeable effect on the OLI SWIR reflectances and the corresponding fire detection performance (Supple- mentary material). Quantitative daytime commission error rates of the algorithm were estimated for 13 different locations showing very distinct levels of fire Fig. 4. Theoretical probability of fire detection calculated for the Landsat-8 algorithm. Day- time (D) curve describes the 50% probability of detection (solid line), bounded by 10% activity (Table 2). Processing of individual scenes included the main (bottom dotted line) and 90% (top dotted line) probability curves. The nighttime fire algorithm and the complementary multi-temporal analysis proce- (N) curve describes the fire area and temperature combination resulting in OLI channel dure described above. The number of usable scenes included in the tem- 2 7 radiance greater than 1 W/(m srμm). poral analysis was dependent on cloud coverage. Argentina and north India were particularly affected by cloud obscuration, with less than products (e.g., S-NPP/VIIRS) (Schroeder et al., 2014), with typical five of the eleven possible scenes acquired in the previous 176 days firewood combustion (950 K mean temperature) requiring ≈ 4 showing cloud-free data. Despite the smaller number of scenes used m2 of effective fire area to achieve greater than 50% chance of in the multi-temporal analysis the algorithm was able to correctly detection. classify active fires, persistent thermal sources, and bright pixels with Our assessment of the nighttime probability of detection curve was zero or negligible commission error rate. based on a simplistic approach involving direct calculation of fire- Six regions of markedly different fire characteristics showed a large emitted radiances from the same set of fire effective area and mean number (N 1000) of pixels detected, namely South Africa, Australia, temperatures described above. We assumed a near-zero and invariant Brazil, Mexico, north India and Iraq. In South Africa and Australia, de- nighttime background radiance, which leaves the simulated fire- tected pixels described large wildfires on each scene. Forest conversion emitted radiance as the only potential source influencing the resulting and pasture maintenance fires were found in Brazil, whereas agricultur- pixel radiance. Similar to the daytime simulations, atmospheric attenu- al fires dominated the scenes in Mexico and north India. The scene in ation effects were estimated using MODTRAN® assuming a U.S. stan- Iraq overlapped with oil fields containing numerous gas flare stations. dard atmospheric profile. Fig. 4 shows the night curve describing the Application of the multi-temporal analysis tests to the locations fire area and temperature combination leading to a channel 7 radiance above showing high land use related biomass burning activity and wild- of 1 W/(m2 sr μm), which would potentially allow for the fire pixel fires resulted in almost no change in pixel classification. The relatively detection. Compared to the daytime curve, nighttime data provide few cases of pixel re-classification were associated with surface coal enhanced response to smaller fires of similar temperature favored by mines (Africa), and with urban structures and scarce vegetation pixels the exclusion of the reflected solar component. For example, the same in other areas. In comparison, application of the multi-temporal analysis typical firewood combustion burning at ≈950 K may be detected tests to the scene in Iraq produced significant changes resulting in when effective fire areas as small as 1 m2 are observed. successful re-assignment of all gas flare pixels to the persistent source The data simulations above did not account for the occurrence of op- class. Commission errors were negligible in all six regions. tically thin or thick clouds and smoke, or the pixel's spatial response The scene in China coincided with urbanized areas characterized by function. Therefore they describe a best-case scenario of fire detection, large factories with highly reflective rooftops. The fire algorithm output which may be subject to partial or complete degradation depending indicated 179 pixels of potential fire activity. Of those, 153 were re- on the prevailing observation conditions. classified by the temporal analysis tests into either persistent source (50%) or bright surface (50%). Of the remaining pixels (26), nine were 6. Algorithm Assessment also coincident with other factories and urban structures indicative of potential false alarms resulting in the highest relative false alarm rate Initial algorithm assessment was based on visual inspection of the of 34.6%. Indonesia was the region showing the second highest commis- output fire pixel coordinates using publicly available data including sion error rate. In that case, 65 pixels were output by the fire algorithm high spatial resolution imagery (e.g., aerial photography, high resolu- and only three were subsequently re-classified by the temporal analysis tion commercial satellite imagery in Google Earth). The fidelity of the tests. Of the 62 remaining pixels, 10 pixels were co-located with an fire detection data was qualitatively assessed by expert image analysts industrial park and therefore considered potential false alarms. Pixels over a wide range of fire conditions, including large/unambiguous and associated with potential false alarms consisted predominantly of single well-documented biomass burning (e.g., boreal fires in Canada, wildfires detections (52%) and small clusters of two adjacent detections (32%). in western U.S.), variable size land management fires (e.g., grassland fires The largest single cluster associated with potential false alarm in Africa, conversion fires in the Amazon, agricultural fires in India), small showed six contiguous pixels overlapping a large industrial complex point sources associated with known gas flare locations (e.g., gas fracking in Indonesia. Overall, global daytime commission error rate was stations in North Dakota/U.S., and oil fields in the middle-east), and other equivalent to 0.2%. verifiable thermal anomalies such as active volcanoes. Complementing the daytime fire algorithm assessment, Fig. 6 shows The Landsat-8 fire algorithm performed well across the wide range the fire detections produced for the nighttime OLI image (path/row of conditions described above. Fig. 5 shows the daytime fire detections 127/217 ascending node) acquired on 04 February 2014 over part of produced for the 2014 King fire in California/U.S., corroborated by North Dakota/U.S. Numerous gas fracking sites are evident in the high same-day fire detections derived from 1-km Terra and Aqua MODIS resolution image of that area, all characterized by small land clearings. Thermal Anomalies (MOD14/MYD14), and 750 m and 375-m S-NPP/ By zooming into individual sites, we could co-locate all fire pixel 216 W. Schroeder et al. / Remote Sensing of Environment 185 (2016) 210–220

Fig. 5. Multi-sensor imaging of the King ﬁre in California/U.S. on 19 September 2014. Active ﬁre detections derived from Terra/MODIS 1 km, Landsat-8/OLI 30 m (path/row 43/33), S-NPP/ VIIRS 375 m and 750 m, and Aqua/MODIS 1 km are marked in yellow. Temporal separation between consecutive images was limited to approximately 1 h.

detections with discernable gas ﬂare stacks on the reference imagery. neighbors, therefore falling below the sensor's minimum resolvable ra- With the exception of one undetected location with a channel 7 radi- diance range. This and other regions inspected showed no evidence of ance of 0.55 W/(m2 sr μm), other fracking sites in the image subset false alarms. As a result, we concluded that nighttime commission er- showed no discernible heat signatures from the background. We as- rors were negligible. Because of the limited number of nighttime scenes sumed the gas ﬂares in those areas could be temporarily out of service available, we could not assess the effect of the multi-temporal analysis or else burning at a much reduced combustion rate relative to their tests in those cases. W. Schroeder et al. / Remote Sensing of Environment 185 (2016) 210–220 217

Table 2 Description of Landsat-8 fire algorithm validation results according to scene location and acquisition date, number of previously acquired scenes used in the multi-temporal analysis, number of pixels output by fire algorithm, number of pixels re-classified as a result of multi-temporal analysis, and estimated commission error rates. The latter describe the percentage of fire algorithm output pixels that were found to be associated with urban pixels (e.g., factories), after exclusion of pixels re-classified by the multi-temporal analysis.

Region Path/row Date (dd/mm/yy) Number of usable scenes Fire algorithm output pixels Pixels re-classiﬁed Commission Error (%) (temporal analysis)

Italy 188/034 23/06/14 10 176 3 1.2 Brazil 227/069 24/06/14 9 355 2 0.0 Brazil 226/069 20/08/14 6 1715 1 0.0 Mexico 026/048 07/05/14 9 1345 6 0.0 USA 023/034 18/05/14 7 27 3 4.8 Australia 105/069 05/08/14 8 3533 0 0.0 Indonesia 127/060 14/06/14 8 65 3 16.1 India (north) 147/040 07/05/14 4 4737 73 0.1 India (south) 143/052 24/03/14 6 444 23 1.7 Argentina 229/085 13/01/14 3 184 1 0.0 Iraq 166/039 19/10/14 9 1400 1289 0.0 South Africa 168/077 15/09/14 10 7154 3 0.0 China 119/041 07/08/14 8 179 153 34.6

6.1. Field verification detection. Processing of the two scenes using the fire algorithm and the temporal analysis tests resulted in two confirmed fire detection Data from two field experiments were also used in support of the fire pixels overlapping each of the experimental plots. Fig. 7 shows the algorithm assessment. On both occasions, a small plot consisting mostly two experimental plots and the corresponding channel 7 radiance of firewood arranged in a rectangular fuel bed was set up in a rural area data subset, plus the resulting fire algorithm output with detection and sampled using ground instrumentation. The first experiment pixels marked in black. The maximum observed radiance on channel 7 (Exp1) was coordinated by the German Aerospace Center (DLR) and im- was equivalent to 27.33 and 24.66 W/(m2 sr μm) in Exp1 and Exp2,re- plemented on 17 August 2013 near Demmin/Germany. The burn plot spectively. Consequently, both fires exceeded the nominal channel 7 measured 11 × 13 m2 and was actively burning during the Landsat-8 saturation value. In comparison, the adjacent fire-free pixels showed a overpass (path/row 194/22) at 10:10 UTC. The approximate mean fire mean radiance of 2.32 and 3.23 W/(m2 sr μm) for Exp1 and Exp2, temperature estimated using a handheld Heitronics thermal sensor respectively. was equivalent to 970 K. The second experiment (Exp2) was coordinat- Additional pixels of elevated radiance on channel 7 bordered the ed by the Brazilian Institute for Space Research (INPE) and implemented two detections on each scene. We attributed those above-background on 19 January 2015 near Cachoeira Paulista/Brazil. A smaller burn plot radiance values to smearing caused by the geometric transformations measuring 3 × 10 m2 was sampled using two uncooled thermal infrared used to convert the OLI Level 1R image (a 2D space consisting of FLIR cameras mounted on tripods positioned adjacent to the fire, plus a image sample/line) to the Level 1 T data (an implicit 3D space consisting dual-band radiometer attached to a 5 m tower at a 45° viewing angle. At of x/y/z projection coordinates), which accounts for the detectors' the time of the Landsat-8 overpass (path/row 218/076) at 12:58 UTC spatial response, observation geometry (line-of-sight), terrain (digital the fire was actively burning with a temperature of 870 ± 153 K. elevation model), and pixel spatial interpolation method (cubic Based on the fire characteristics describing the experimental plots convolution) used to produce the output resampled data in Universal above and the theoretical detection envelope calculated for the Transverse Mercator (UTM) projection (U.S. Department of the Landsat-8 fire algorithm in Section 5, both sites had high probability of Interior – U.S. Geological Survey, 2013). Stray light effects may also contribute to the observed smearing (Morfitt et al., 2015). Such data artifact was systematically observed over fires corresponding to small/ point sources such as gas flares occurring in both day and nighttime images. Nonetheless we expect smearing to equally affect larger fires. In fact, visual inspection of training and validation data frequently showed pixels of intermediate (above-background) radiance values adjacent to pixels classified as fires. Those locations were classified as fire-free pixels by our detection methodology and not considered omission errors. Additional validation analyses using high resolution (≈1m) reference fire data are needed to further corroborate the current results.

7. Conclusions

The biomass burning science and data user community have for many years relied on coarser spatial resolution (≥1 km) satellite data to map and monitor fire-affected areas. Current operational satellite- based active fire detection systems provide routine observations of global fire activity, serving as input to a wide range of science applications, air quality monitoring, and strategic fire management. However, the lack of spatial fidelity typical of those satellite active fire products remains a major limitation preventing their application in landscape Fig. 6. Subset of the fire detection algorithm output (red polygons) derived from the 04 analyses and tactical fire management. February 2014 Landsat-8/OLI nighttime scene path/row 127/217 (ascending node) over The ever-growing number of high and moderate spatial resolution part of North Dakota/U.S. Background image shows 1 m resolution true-color composite aerial photography data acquired in July–August 2014 by the U.S. Department of Agricul- earth observation systems, including Landsat-class instruments, is ture National Agriculture Imagery Program. poised to change that reality. The launch in 2013 of Landsat-8 added 218 W. Schroeder et al. / Remote Sensing of Environment 185 (2016) 210–220

Fig. 7. Experimental fires used in support of Landsat-8 fire algorithm verification. The two rows show photos of the sites (left panels), and the OLI channel 7 image subset of the area with fire pixels marked in black (right panels). Top panels describe the 11 × 13 m2 plot near Demmin/Germany (53.9276°N 13.0587°E) on 17 August 2013; bottom panels describe the 3 × 10 m2 plot near Cachoeira Paulista/Brazil (22.6868°S44.9844°W) on 19 January 2015.

to the existing network of science-quality moderate spatial resolution number. In addition, visual inspection of the data corroborated by field sensors; the ongoing planning and designing of the follow up Landsat- verification indicated potential smearing of the fire-emitted radiance 9 system could further extend the Landsat time series into the next across adjacent pixels. We attributed the smearing effect to the resam- decade. Most importantly, the comparatively recent shift towards an pling procedure involved in the production of the Level 1 T data files, open data policy helped catapult the use of Landsat data among various and to residual stray light effects. science applications, including biomass burning. Complemented by Despite those unique input data features, our assessment of the out- other missions of similar scope (e.g., ESA/Sentinel-2), these emerging put fire classification data indicated good overall quality across a wide orbital systems have the potential to transform the way satellite data range of fire conditions observed in both day and nighttime imagery. are used in support of fire science and management applications. Collec- Small and large-size biomass burning fires were successfully resolved, tively, the integration of all available Landsat-class instruments in the along with gas flares and active volcanoes. The high quality of the near future could provide satellite-based spatially refined fire detection Level 1 T geolocation data provided improved co-location of pixels data at a temporal resolution frequent enough to support operational acquired on multiple dates enabling the post-processing of the fire algo- tactical fire management, adding to fire incident teams' limited array rithm output for persistent heat anomalies, and separation of potential of spatially-explicit fire perimeter information. false alarms. Overall commission errors were low, averaging 0.2% Building on these new capabilities, we presented a new active fire globally. Theoretical calculation of fire detection envelopes suggested algorithm using Landsat-8/OLI day and nighttime input data. The meth- significant improvement in performance compared to existing coarser odology expands on previous algorithms proposed for ASTER and spatial resolution fire products. Field verification data corroborated Landsat-7/ETM+, incorporating additional visible and near-infrared our theoretical assessment. channel data and a multi-temporal analysis scheme. Detailed analyses The spatially-refined active fire detection data provided by Landsat- of the input OLI data indicated frequent saturation of fire-affected pixels class sensors create new opportunities and challenges for the user and on the fire-sensitive SWIR channel 7, and to a lesser extent of the NIR science communities. Thanks to the improved fire line resolving capa- channel 6 data. Under more extreme conditions involving large or bilities, pixel-based analyses could be replaced by cluster-based applica- high intensity fires, channel 7 saturation was greatly exceeded resulting tion of the data where contiguous pixels are labeled accordingly and in radiometric artifacts associated with the folding of the pixel's digital used to describe individual fire events. Such data application could W. Schroeder et al. / Remote Sensing of Environment 185 (2016) 210–220 219 also lead to revised product assessment metrics describing omission Davies, D. K., Ilavajhala, S., Wong, M. M., & Justice, C. O. (2009). Fire Information for Re- source Management System: Archiving and distributing MODIS active fire data. and commission error rates based on clustering analyses. Detailed IEEE Transactions on Geoscience and Remote Sensing, 47,72–79. validation of these data remains a challenge as reference fire data Dwyer, J., Pinnock, S., Grégoire, J. -M., & Pereira, J. M. C. (2000). Global spatial and tempo- sources are scarce and costly to acquire, currently demanding ral distribution of vegetation fire as determined from satellite observations. International Journal of Remote Sensing, 21,1289–1302. exhaustive image interpretation by expert image analysts. Future al- Elvidge, C. D., Ziskin, D., Baugh, K. E., Tuttle, B. T., Ghosh, T., Pack, D. W., ... Zhizhin, M. gorithm assessment should build on new complementary data sets (2009). A fifteen year record of global natural gas flaring derived from satellite (e.g., high resolution urban maps to identify potential false alarms data. Energies, 2,595–622. fi over highly reflective buildings) and improved fire detection data Elvidge, C. D., Zhizhin, M., Hsu, F. -C., & Baugh, K. E. (2013). VIIRS night re: Satellite py- rometry at night. Remote Sensing, 5, 4423–4449. simulation corroborated by field data acquired using new lower Flynn, L. P., Harris, A. J. L., & Wright, R. (2001). Improved identification of volcanic features cost technologies (e.g., drones). using Landsat 7 ETM+. Remote Sensing of Environment, 78,180–193. The increased availability of Landsat-class active fire data is fostering Francis, P. W., & Rothery, D. A. (1987). Using the Landsat Thematic Mapper to detect and monitor active volcanoes: An example from Lascar volcano, northern Chile. Geology, the development of new applications focused on landscape-scale analy- 15,614–617. ses, including fire growth assessment, fire data assimilation into Freeman, J. P., Stohlgren, T. J., Hunter, M. E., Omi, P. N., Martinson, E. J., Chong, G. W., & cutting-edge coupled weather-fire models, and tactical fire manage- Brown, C. S. (2007). Rapid assessment of postfire plant invasions in coniferous forests of the Western United States. Ecological Applications, 17,1656–1665. ment, among others. In order to further maximize time-sensitive disas- Giglio, L., Descloitres, J., Justice, C. O., & Kaufman, Y. J. (2003). An enhanced contextual fire ter monitoring applications, mission requirements should enable near detection algorithm for MODIS. Remote Sensing of Environment, 87,273–282. real-time processing and distribution of the data. The global sampling Giglio, L., Csiszar, I., & Justice, C. O. (2006). Global distribution and seasonality of fires as observed with the Terra and Aqua Moderate Resolution Imaging Spectroradiometer strategy should also include comprehensive and systematic nighttime (MODIS) sensors. Journal of Geophysical Research, 111. http://dx.doi.org/10.1029/ acquisitions of shortwave and thermal infrared data that can be used 2005JG000142. to more frequently track the progress of thermal anomalies. Giglio, L., Csiszar, I., Restás, Á., Morisete, J. T., Schroeder, W., Morton, D., & Justice, C. O. (2008). Active fire detection and characterization with the Advanced Spaceborne The proposed algorithm was implemented at the USDA Remote Thermal Emission and Reflection Radiometer (ASTER). Remote Sensing of Sensing Applications Center in Salt Lake City/UT, where it serves the Environment, 112,3055–3063. U.S. fire management community. This and other Landsat-class active Giglio, L., Loboda, T. V., Roy, D. P., Quayle, B., & Justice, C. O. (2009). An active-fire based fire algorithms will continue to be developed and further refined. Future burned area mapping for the MODIS sensor. Remote Sensing of Environment, 113, 408–420. development of a fire disturbance essential climate variable will be pur- Hantson, S., Padilla, M., Corti, D., & Chuvieco, E. (2013). Strengths and weaknesses of sued, complementing other satellite fire data sets in support of land- MODIS hotspots to characterize global fire occurrence. Remote Sensing of – science applications. Environment, 131,152 159. Higgins, S. I., Bond, W. J., February, E. C., Bronn, A., Euston-Brown, D. I. W., Enslin, B., ... Supplementary data to this article can be found online at http://dx. Trollope, W. S. W. (2007). Effects of four decades of fire manipulation on woody veg- doi.org/10.1016/j.rse.2015.08.032. etation structure in savanna. Ecological Applications, 88,1119–1125. Ichoku, C., & Kaufman, Y. J. (2005). A method to derive smoke emission rates from MODIS fire radiative energy measurements. IEEE Transactions on Geoscience and Remote Acknowledgments Sensing, 43,2636–2649. Ichoku, C., Giglio, L., Wooster, M. J., & Remer, L. A. (2008). Global characterization of biomass-burning patterns using satellite measurements of fire radiative energy. We thank DLR, Erik Borg, and Doris Klein for hosting and sharing Remote Sensing of Environment, 112, 2950–2962. data from the small biomass burning experiment in Demmin/ Irons, J. R., Dwyer, J. L., & Barsi, J. A. (2012). the next landsat satellite: the landsat data con- tinuity mission. Remote Sensing of Environment, 112,11–21. Germany on 17 August 2013. We also thank INPE for hosting the small Kaiser, J. W., Heil, A., Andreae, M. O., Benedetti, A., Chubarova, N., Jones, L., ... van der Werf, biomass burning experiment on 19 January 2015 in Cachoeira G. (2012). Biomass burning emissions estimated with a global fire assimilation sys- Paulista/Brazil, and André Lima for sharing field data. Funding support tem based on observed fire radiative power. Biogeosciences, 9,527–554. Kasischke, E. S., Hewson, J. H., Stocks, B., van der Werf, G., & Randerson, J. (2003). The use for this study was provided by the NASA Applied Sciences Program of ATSR active fire counts for estimating relative patterns of biomass burning — A (grant #NNX12AQ87G), the USDA Forest Service Remote Sensing Appli- study from boreal forest region. Geophysical Research Letters, 30. http://dx.doi.org/ cations Center, and the GOES-R Algorithm Working Group at NOAA's 10.1029/2003GL017859. National Environmental Satellite, Data, and Information Service. We Kaufman, Y. J., Wald, A. E., Remer, L. A., Gao, B. -C., Li, R. -R., & Flynn, L. (1997). The MODIS 2.1 μm channel — Correlation with visible reflectance for use in remote sensing of thank four anonymous reviewers for their comments. aerosol. IEEE Transactions on Geoscience and Remote Sensing, 35,1286–1298. Larsen, I. J., MacDonald, L. H., Brown, E., Rough, D., Welsh, M. J., Pietraszek, J. H., ... Schaffrath, K. (2009). Causes of post-fire runoff and erosion: water repellency, References cover or soil sealing? Soil Science Society of America Journal, 73,1393–1407. Liu, H., Randerson, J. T., Lindfors, J., & Chapin, F. S., III (2005). Changes in the surface energy Anejionu, O. C. D., Blackburn, G. A., & Whyatt, J. D. (2014). Satellite survey of gas flares: budget after fire in boreal ecosystems of interior Alaska: An annual perspective. Development and application of a Landsat-based technique in the Niger Delta. Journal of Geophysical Research-Atmospheres. http://dx.doi.org/10.1029/2004JD International Journal of Remote Sensing, 35(5), 1900–1925. 005158. Asner, G. P., & Lobell, D. B. (2000). A biogeophysical approach for automated SWIR Loboda, T. V., & Csizar, I. (2007). Reconstruction of fire spread within wildland fire events unmixing of soils and vegetation. Remote Sensing of Environment, 74,99–112. in Northern Eurasia from the MODIS active fire product. Global and Planetary Change, Bowman, D. M. J. S., Balch, J. K., Artaxo, P., Bond, W. J., Carlson, J. M., Cochrane, M. A., ... 56,257–273. Pyne, S. J. (2009). Fire in the earth system. Science, 324,481–484. Loboda, T. V., O'Neal, K., & Csizar, I. (2007). Regionally adaptable dNBR-based algorithm Casadio, S., Arino, O., & Serpe, D. (2012). Gas flaring monitoring from space using the for burned area mapping from MODIS data. Remote Sensing of Environment, 109, ATSR instrument series. Remote Sensing of Environment, 116,239–249. 429–442. Coen, J. L., & Schroeder, W. (2013). Use of spatially refined satellite remote sensing fire de- Mell, W. E., Manzello, S. L., Maranghides, A., Butry, D., & Rehm, R. G. (2010). The wildland- tection data to initialize and evaluate coupled weather-wildfire growth model simu- urban interface fire problem — Current approaches and research needs. International lations. Geophysical Research Letters. http://dx.doi.org/10.1002/2013GL057868. Journal of Wildland Fire, 19,238–351. Coen, J. L., & Schroeder, W. (2015). The high park fire: Coupled weather-wildland fire Morfitt, R., Barsi, J., Levy, R., Markham, B., Micijevic, E., Ong, L., ... Vanderwerff, K. (2015). model simulation of a windstorm-driven wildfire in Colorado's front range. Journal Landsat-8 Operational Land Imager (OLI) radiometric performance on-orbit. Remote of Geophysical Research-Atmospheres, 120. http://dx.doi.org/10.1002/2014JD021993. Sensing, 7,2208–2237. Csiszar, I. A., & Schroeder, W. (2008). Short-term observations of the temporal develop- Morisette, J. T., Giglio, L., Csiszar, I., & Justice, C. O. (2005). Validation of the MODIS active ment of active fires from consecutive same-day ETM+ and ASTER imagery in the fire product over Southern Africa with ASTER data. International Journal of Remote Amazon: Implications for active fire product validation. IEEE Journal of Selected Sensing, 26(19), 4239–4264. Topics in Applied Earth Observations and Remote Sensing, 1,248–253. Oliva, P., & Schroeder, W. (2015). Assessment of VIIRS 375 m active fire detection product Csiszar, I. A., Morisette, J. T., & Giglio, L. (2006). Validation of active fire detection from for direct burned area mapping. Remote Sensing of Environment, 160,144–155. moderate-resolution satellite sensors: The MODIS example in Northern Eurasia. Oliveras, I., Meirelles, S. T., Hirakuri, V. L., Freitas, C. R., Miranda, H. S., & Pivello, V. R. IEEE Transactions on Geoscience and Remote Sensing, 44(7), 1757–1764. (2012). Effects of fire regimes on herbaceous biomass and nutrient dynamics in the Csiszar, I., Schroeder, W., Giglio, L., Ellicott, E., Vadrevu, K., Justice, C. O., & Wind, B. (2014). Brazilia savanna. International Journal of Wildland Fire, 22,368–380. Active fires from the Suomi NPP Visible Infrared Imaging Radiometer Suite: Product Oppenheimer, C. (1991). Lava flow cooling estimated from Landsat Thematic Mapper status and first evaluation results. Journal of Geophysical Research-Atmospheres,119. data: The Lonquimay Eruption (Chile, 1989). Journal of Geophysical Research, 96. http://dx.doi.org/10.1002/2013JD020453. http://dx.doi.org/10.1029/91JB01902. 220 W. Schroeder et al. / Remote Sensing of Environment 185 (2016) 210–220

Pozo, D., Olmo, F. J., & Alados-Arboledas, L. (1997). Fire detection and growth monitoring U.S. Department of the Interior – U.S. Geological Survey (2013). LDCM CAL/VAL Algorithm using a multitemporal technique on AVHRR mid-infrared and thermal channels. Description Document, version 3.0. Digital document available at: http://landsat. Remote Sensing of Environment, 60,111–120. usgs.gov/documents/LDCM_CVT_ADD.pdf (last accessed 22 June 2015) Roy, D., Wulder, M. A., Loveland, T. R., Woodcock, C. E., Allen, R. G., Anderson, M. C., ... Zhu, Van der Werf, G. R., Randerson, J. T., Giglio, L., Collatz, G. J., Mu, M., Kasibhatla, P. S., ... van Z. (2014). Landsat-8: Science and product vision for terrestrial global change re- Leeuwen, T. T. (2010). Global fire emissions and the contribution of deforestation, sa- search. Remote Sensing of Environment, 145,154–172. vanna, forest, agricultural, and peat fires (1997–2009). Atmospheric Chemistry and Schoennagel, T., Nelson, C. R., Theobald, D. M., Carnwath, G. C., & Chapman, T. B. (2009). Physics, 10,11707–11735. Implementation of National Fire Plan treatments near the wildland-urban interface Vermote, E., Ellicott, E., Dubovik, O., Lapyonok, T., Chin, M., Giglio, L., & Roberts, G. J. in the western United States. Proceedings of the National Academy of Sciences, 106, (2009). An approach to estimate global biomass burning emissions of organic and 10706–10711. black carbon from MODIS fire radiative power. Journal of Geophysical Research, 114. Schroeder, W., Prins, E., Giglio, L., Csiszar, I., Schmidt, C., Morisette, J., & Morton, D. http://dx.doi.org/10.1029/2008JD011188. (2008a). Validation of GOES and MODIS active fire detection products using ASTER Wiedinmyer, C., Quayle, B., Geron, C., Belote, A., McKenzie, D., Zhang, X., ... Wynne, K. K. and ETM+ data. Remote Sensing of Environment, 112,2711–2726. (2006). Estimating emissions from fires in North America for air quality modeling. Schroeder, W., Ruminski, M., Csiszar, I., Giglio, L., Prins, E., Schmidt, C., & Morisette, J. Atmospheric Environment, 40, 3419–3432. (2008b). Validation analyses of an operational fire monitoring product: The hazard mapping system. International Journal of Remote Sensing, 29,6059–6066. Schroeder, W., Oliva, P., Giglio, L., & Csiszar, I. (2014). The new VIIRS 375 m active fire detection data product: Algorithm description and initial assessment. Remote Sensing of Environment, 143,85–96.